environment-life-review-forge - SKILL.md Agent Skill

name: environment-life-review-forge description: Adapts evidence synthesis workflows for environmental, ecological, biomedical, and life-science questions. Use for PECO/PICO frameworks, exposure-outcome reviews, ecological heterogeneity, dose-response evidence, risk-of-bias planning, environmental indicators, NDVI or vegetation-index models, partial least squares regression, PLS VIP audits, ecosystem-service relationships, ESR synergy/trade-off mapping, interpretable machine learning, GWR/XGBoost spatial modeling, threshold-oriented ecological management, optimal interval identification, air pollution crop-yield models, ozone/aerosol food-security co-benefits, SIF-based crop productivity, soil biodiversity, aridity gradients, ecosystem stability, climate-stress moderation, soil fauna meta-analysis, ant-mediated carbon cycling, SOC and CO2 dual-outcome synthesis, organism/tissue/time-scale coding, wetland methane scaling, small-patch geospatial upscaling, cryosphere or permafrost evidence products, near-surface ground ice mapping, geospatial environmental maps, spatial autocorrelation audits, map uncertainty audits, food-system environmental nexus reviews, food-waste geospatial ML forecasting, agroecosystem nutrient meta-analysis, crop-yield machine-learning prediction, environmental causal machine learning, double machine learning, spatial DML, urban heat causal inference, interpretable exposure-response modeling, genotype-environment modeling, agricultural irrigation optimization, brackish-water irrigation, GAM or NSGA-II trade-off modeling, environmental scenario modeling, land-use optimization, Pareto-frontier trade-off synthesis, policy trade-off synthesis, system-hub environmental variables, nitrogen-SDG coupling, safe-boundary synthesis, hotspot-layer identification, and domain-specific systematic review protocols.

Environment Life Review Forge

Use this skill for environmental, ecological, biomedical, and life-science systematic reviews where exposure, organism/population, outcome, context, and study design need careful domain adaptation.

Core Principle

Domain structure matters. The same effect-size workflow may be misleading if exposure windows, species, tissues, endpoints, geography, or measurement platforms are not comparable.

Intake

Identify:

domain: environment, ecology, toxicology, epidemiology, life science, molecular biology, public health;
framework: PECO or PICO;
population or organism;
exposure or intervention;
comparator;
outcomes/endpoints;
study design;
spatial and temporal scale;
ecosystem-service set, pairwise relationship definition, synergy/trade-off coding, and threshold-management target, if ecosystem-service relationships are in scope;
pollutant exposure, crop outcome, food-security endpoint, counterfactual air-quality target, and crop-calorie translation, if air-quality food-security modeling is in scope;
biodiversity dimension, stability metric, climate-stress gradient, and moderation/interaction target, if biodiversity-stability evidence is in scope;
spatial unit and aggregation boundary, if geospatial prediction is in scope;
minimum mapping unit and detection threshold, if small-patch systems are in scope;
target map variable, spatial resolution, observation inventory, predictor stack, spatial autocorrelation plan, and uncertainty layer, if an environmental map product is in scope;
measurement method;
bidirectional pathways, if impacts and feedbacks are both in scope;
expected heterogeneity.

Load:

references/environmental-life-science.md for domain heterogeneity.
references/cee-alignment.md for environmental evidence standards.
references/pls-vip-environmental-indicators.md for NDVI, vegetation, soil, climate, ecological indicator, PLS regression, and VIP interpretation audits.
references/ecosystem-service-threshold-ml.md for ecosystem-service relationship mapping, GWR-plus-ML workflows, nonlinear driver interpretation, threshold/optimal-interval identification, and spatial management translation.
references/air-quality-food-security.md for ozone, aerosol, SIF, crop yield, crop-calorie, counterfactual air-quality targets, and food-security co-benefit modeling.
references/soil-biodiversity-aridity-stability.md for soil biodiversity, aridity gradients, ecosystem stability, climate-stress moderation, and biodiversity-function buffering claims.
references/ant-soil-carbon-meta.md for soil-fauna meta-analysis, ecosystem-engineer effects on SOC stock and CO2 flux, trait-mediated moderators, and climate-context extraction.
references/small-wetland-methane-scaling.md for wetland methane, small water bodies, fine-resolution remote sensing, and scale-sensitive upscaling.
references/cryosphere-ground-ice-mapping.md for permafrost, near-surface ground ice, borehole observations, geospatial predictors, ensemble machine learning, spatial autocorrelation, prediction intervals, and public map-data audits.
references/agroecosystem-nutrient-meta-analysis.md for crop yield, soil organic carbon, fertilizer, amendment, and nutrient-management meta-analyses.
references/agricultural-ml-yield-prediction.md for crop-yield prediction studies integrating meteorological, breeding, genomic, remote-sensing, or field-trial data.
references/agricultural-irrigation-optimization.md for brackish-water irrigation, water-salt-yield-emission trade-offs, GAM nonlinear response modeling, NSGA-II optimization, and decision ranges such as ECw management windows.
references/environmental-causal-ml.md for environmental causal machine learning studies using DML, CATE, AutoML, SHAP/PDP-style interpretation, high-dimensional pollutant exposure data, socioeconomic covariates, ARGs, drinking-water safety, or One Health outcomes.
references/food-system-bidirectional-nexus.md for food-system reviews linking environmental pressures, feedbacks, trade, diets, crops, livestock, and aquatic foods.
references/food-waste-geospatial-ml.md for county, city, supply-chain, or market-level food-waste forecasting with geospatial analytics and machine learning.
references/environmental-scenario-synthesis.md when a review builds a literature-derived database, machine-learning/spatial model, or policy scenario simulation.
references/land-use-optimization-tradeoffs.md when a study uses multiobjective optimization, Pareto frontiers, land-use allocation, or food-water-carbon trade-off modeling.
references/system-hub-policy-synthesis.md when a paper uses one focal variable, such as nitrogen, carbon, water, phosphorus, air pollution, or biodiversity pressure, to connect multiple environmental, production, health, or policy outcomes under a boundary or scenario framework.

Workflow

Build PECO/PICO.
Define exposure or intervention precisely.
Define outcome families and measurement units.
Specify eligible designs.
Identify heterogeneity sources.
Plan risk-of-bias or study quality appraisal.
Plan grey-literature and supplementary search if relevant.
Decide narrative, evidence map, or meta-analysis.
Build domain-specific extraction table.

Use templates/peco-framework.md. Use templates/pls-vip-environmental-audit.md for PLS/VIP environmental indicator studies. Use templates/ecosystem-service-threshold-audit.md and templates/ecosystem-service-threshold-schema.csv for ecosystem-service relationship threshold-management studies. Use templates/air-quality-food-security-audit.md and templates/air-quality-food-security-schema.csv for air-pollution, crop-yield, SIF, and food-security co-benefit studies. Use templates/biodiversity-stability-climate-stress-audit.md and templates/biodiversity-stability-climate-stress-schema.csv for soil biodiversity, aridity, ecosystem-stability, and climate-stress moderation studies. Use templates/soil-fauna-carbon-meta-audit.md, templates/soil-fauna-carbon-schema.csv, and templates/soil-fauna-carbon-method-stack-schema.csv for ant, termite, earthworm, or other soil-fauna meta-analyses that separate SOC stock, CO2 flux, organic-matter stability outcomes, and method-stack choices such as multilevel meta-analysis, random forest, and PLS-PM/path modeling. Use templates/wetland-methane-scale-audit.md and templates/wetland-methane-geospatial-schema.csv for small-wetland methane and scale-sensitive upscaling studies. Use templates/cryosphere-ground-ice-map-audit.md and templates/cryosphere-map-validation-schema.csv for permafrost, near-surface ground ice, and other cryosphere map products. Use templates/food-environment-bidirectional-audit.md and templates/food-environment-pressure-schema.csv for food-system nexus reviews. Use templates/food-waste-forecast-audit.md and templates/food-waste-geospatial-feature-schema.csv for geospatial food-waste forecasting. Use templates/dual-outcome-meta-audit.md and templates/nutrient-meta-extraction-schema.csv for agroecosystem nutrient meta-analysis. Use templates/nutrient-meta-reproducibility-ledger.csv when a nutrient meta-analysis provides Zenodo/OSF/GitHub data and code. Use templates/nutrient-meta-dataset-schema.csv and templates/nutrient-meta-r-workflow-blueprint.csv when designing data tables and R scripts for nutrient meta-analysis. Use templates/ml-yield-prediction-audit.md and templates/ml-yield-feature-schema.csv for agricultural ML yield-prediction studies. Use templates/environmental-causal-ml-audit.md and templates/environmental-causal-ml-feature-schema.csv for environmental causal ML studies. Use templates/irrigation-optimization-audit.md and templates/irrigation-optimization-schema.csv for brackish-water irrigation and water-salt-yield-emission optimization studies. Use templates/scenario-model-audit.md and templates/policy-scenario-matrix.csv for scenario-model evidence synthesis. Use templates/pareto-frontier-audit.md and templates/multi-objective-tradeoff-schema.csv for multiobjective optimization and land-use trade-off studies. Use templates/system-hub-variable-audit.md and templates/system-hub-variable-schema.csv for papers that organize evidence around a system hub variable, safe boundary, hotspot layer, co-benefit structure, or policy-portfolio translation.

Output Modes

PECO Protocol Memo

Population:
Exposure:
Comparator:
Outcome families:
Eligible designs:
Heterogeneity:
Risk of bias:
Synthesis plan:

Domain Extraction Plan

Include:

species/population;
site/geography;
exposure dose/intensity;
duration/window;
endpoint;
assay/measurement method;
confounders;
study quality.

PLS/VIP Environmental Indicator Audit

Use this mode when a study links NDVI, vegetation productivity, soil quality, biodiversity, ecosystem-service, pollutant, climate, or hydrological indicators to multiple correlated predictors using partial least squares regression and VIP rankings.

Include:

outcome indicator and unit;
predictor families and expected ecological meaning;
sample size, time span, site count, and clustering;
missing-data and scaling decisions;
component-count selection rule;
cross-validation design and whether it respects time, site, or spatial grouping;
RMSEP, R2/Q2, residual checks, and influential observations;
VIP table, VIP threshold, and coefficient direction;
whether VIP is interpreted as predictive importance, mechanism, or causal effect;
robustness checks such as alternative component count, leave-one-year/site-out validation, or baseline regression comparison.

Ecosystem-Service Threshold ML Audit

Use this mode when a study maps ecosystem services or ecosystem-service relationships, then uses interpretable machine learning, GWR, GAM, XGBoost, GBDT, SHAP, PDP, ICE, ALE, or response-curve overlays to identify management thresholds or optimal driver intervals.

Include:

ecosystem services and pairwise ESR definitions;
spatial unit, resolution, years, and service-assessment models;
relationship coding: synergy, trade-off, co-benefit, competition, or probability of comprehensive synergy;
driver families: climate, topography, landscape, land use, human pressure, accessibility, and policy;
spatial heterogeneity model such as GWR and its bandwidth/kernel choices;
ML model set, tuning, validation, and spatial leakage checks;
interpretability method and whether thresholds come from PDP, SHAP dependence, ALE, response-curve superposition, or another method;
optimal interval definition, probability target, and uncertainty;
zoning or management translation and whether it is descriptive, predictive, or causal;
robustness checks against alternative service models, spatial resolution, model family, and threshold rule.

Air Quality Food-Security Audit

Use this mode when a study estimates how ozone, PM2.5, aerosol optical depth, diffuse radiation, SIF, temperature, or related atmospheric conditions affect crop productivity, crop yield, calories, self-sufficiency, or food-security indicators.

Include:

crops, regions, years, and spatial unit;
pollutant exposure metrics, such as AOT40, peak-season ozone, PM2.5, AOD, or aerosol loading;
crop outcome: yield, SIF, productivity proxy, calorie output, or supply-demand balance;
statistical model, flexible functional form, crop fixed effects, spatial effects, weather controls, and trend controls;
counterfactual air-quality targets and whether targets are policy-based;
nonlinear and synergistic pollutant-response claims;
validation against observed yields or independent crop productivity data;
translation from yield to calories, self-sufficiency, imports, or food security;
data/code availability and uncertainty in source data, crop area, crop-calorie conversion, and counterfactual scenarios.

Biodiversity Stability Climate-Stress Audit

Use this mode when a study evaluates whether biodiversity, soil biodiversity, microbial diversity, functional diversity, or community composition supports ecosystem stability under aridity, drought, warming, land-use stress, or other climate gradients.

Include:

biodiversity dimension and measurement platform;
stability endpoint: temporal stability, resistance, resilience, multifunctionality, productivity variability, or service stability;
climate-stress gradient and exposure window;
ecosystem type, spatial domain, sampling design, and temporal depth;
interaction or moderation model: whether aridity weakens, strengthens, or changes the biodiversity-stability relationship;
controls for productivity, soil, climate, land use, management, and spatial dependence;
mechanism evidence from microbes, soil nutrients, plant traits, or food-web structure;
uncertainty, threshold, nonlinear, and subgroup evidence;
whether claims are observational associations, experiments, manipulations, or causal estimates.

Soil Fauna Carbon Meta-Analysis Audit

Use this mode when a study synthesizes how ants or other soil fauna affect soil carbon storage, carbon fluxes, organic matter turnover, or stability across ecosystems.

Include:

focal fauna and functional traits;
stock versus flux versus stability outcome separation;
effect-size family and percent-change interpretation;
climate, latitude, baseline SOC, or ecosystem moderators;
dependence plan for multiple endpoints per study;
whether driver ranking uses random forest or another ML method;
whether indirect pathways are organized with PLS-PM, SEM, or another path model;
whether "more storage" and "more emissions" are both present and how the paper interprets that pattern.

Small-Wetland Methane Scaling Audit

Use this mode when a study estimates emissions from small wetlands, ponds, small water bodies, or patchy ecosystems where spatial resolution and minimum mapping unit change the global or regional budget.

Include:

wetland or water-body size class;
mapping resolution and minimum detectable area;
forested/non-forested domain boundary;
wetland inventory or remote-sensing product;
flux model or emission-factor source;
annual trend window;
contribution to total emissions;
uncertainty, double-counting, and omission risks;
implications for methane budgets and restoration policy.

Cryosphere Ground-Ice Map Audit

Use this mode when a study creates or reuses a spatial map product for permafrost, near-surface ground ice, volumetric ice content, active-layer properties, thermokarst susceptibility, or related cryosphere hazards.

Include:

target map variable and depth convention;
spatial domain, permafrost mask, map year/window, resolution, and grid definition;
field observation type, count, spatial distribution, and measurement comparability;
predictor stack: substrate, hydrology, topography, geology, paleoclimate, modern climate, remote sensing, and vegetation;
model families, ensemble strategy, calibration data, and simulation count;
validation design, including independent or spatially blocked validation;
spatial autocorrelation, sampling bias, and extrapolation checks;
accuracy, bias, RMSE, R-squared, prediction interval, and uncertainty maps;
storage, extent, hazard, infrastructure, climate, hydrology, and ecosystem interpretations;
data DOI, code availability, and versioning of map products.

Agroecosystem Nutrient Meta-Analysis

Use this mode for fertilizer, manure, compost, liming, biochar, and nutrient-management reviews with crop, soil, emission, or microbial outcomes.

Include:

intervention nutrient form and rate;
comparator nutrient background;
crop or ecosystem;
soil baseline status;
climate and geography;
experiment duration;
yield endpoint;
soil-carbon or soil-health endpoint;
response-ratio or percent-change metric;
moderator plan;
dual-outcome trade-off interpretation.

Food-System Bidirectional Nexus Review

Use this mode when food production affects environmental pressures and those environmental changes feed back onto food production.

Include:

food-system scope and commodity groups;
pressure domains: nutrients, water, biodiversity, climate, land use, soil, pollutants;
feedback receptors: crops, livestock, blue foods, food security;
trade or displacement pathways;
supply-side strategies;
demand-side strategies;
circularity or redesign strategies;
regional vulnerability and equity concerns;
evidence certainty and data-source limits.

System-Hub Variable Audit

Use this mode when a paper treats nitrogen, carbon, water, phosphorus, potassium, air pollution, biodiversity pressure, or another focal variable as a systems connector linking environmental burdens, productivity, health, food security, ecosystem function, or policy goals.

Include:

focal system variable and why it is treated as a hub;
coupled outcome families and which are measured versus translated;
boundary, threshold, or target definition;
sectors, regions, or layers included;
hotspot or hidden-layer logic;
co-benefit and trade-off pattern;
scenario or policy-portfolio structure;
uncertainty in the translation from biophysical result to policy claim;
cross-domain lesson that could transfer to other nexus variables.

Food-Waste Geospatial ML Forecast

Use this mode when a study predicts food waste or organic waste across counties, cities, markets, facilities, or supply-chain regions.

Include:

waste stream and supply-chain boundary;
spatial unit and resolution;
target estimate source and measurement limits;
demographic, retail, expenditure, income, policy, and access predictors;
model families and hyperparameter tuning;
spatial validation and leakage checks;
feature importance and uncertainty;
logistics, facility siting, circular bioeconomy, or waste-to-resource interpretation;
limits on translating predicted waste into collectible feedstock.

Crop-Yield ML Prediction Audit

Use this mode when a study predicts crop yield from weather, genotype, breeding value, trial, remote-sensing, soil, or management inputs.

Include:

crop, region, years, sites, genotypes, and trial design;
input data families and temporal aggregation;
target yield definition and unit;
model families compared;
hyperparameter tuning and validation split;
leakage risks across site, year, genotype, and repeated trials;
performance metrics;
interpretability method;
extrapolation and future-climate assumptions;
reproducibility and data-access status.

Environmental Causal ML Audit

Use this mode when a study claims causal effects from high-dimensional environmental, pollutant, socioeconomic, microbial, omics, health, or infrastructure data using DML, causal forests, meta-learners, AutoML-plus-causal-inference, or interpretable ML.

Include:

unit of observation and sampling frame;
treatment/exposure definitions;
outcome endpoint and measurement platform;
covariate timing and confounder set;
predictive model and causal estimator;
nuisance-model, cross-fitting, and overlap checks;
interpretability method and whether it is causal or predictive;
sensitivity to hidden confounding, spatial dependence, and multiple testing;
policy or risk-assessment claim and limits.

For urban heat DML, also include:

treatment or exposure such as heatwave event, green space, impervious surface, shade, cool roof, ventilation, or adaptation policy;
outcome such as LST, SUHI, air temperature, heat exposure, nighttime lights, activity, energy use, or health;
remote-sensing, morphology, meteorology, population, income, activity, and infrastructure confounders;
spatial autocorrelation, spillover, clustered inference, and panel/event-study structure;
whether the claim is average effect, heterogeneous effect, or mediation.

Irrigation Salinity Optimization Audit

Use this mode when a study combines field trials, statistical response curves, and multiobjective optimization to choose irrigation or nutrient-management windows.

Include:

crop, region, irrigation method, salinity or nutrient treatments;
response variables: yield, quality, soil properties, GHG emissions, water use, salt balance;
model used to estimate nonlinear responses, such as GAM;
optimizer used to define compromise ranges, such as NSGA-II;
objectives, constraints, units, and direction of improvement;
decision range and whether it is site-specific;
risks from short time windows, salinity accumulation, unmeasured costs, and extrapolation.

Environmental Scenario Audit

Use this mode when papers combine study extraction, model prediction, spatial extrapolation, and policy scenarios.

Include:

evidence database scope;
target variable and units;
model family and validation plan;
spatial/temporal extrapolation assumptions;
scenario levers;
management-quality assumptions;
uncertainty and sensitivity checks;
policy claim and limits.

Pareto-Frontier Trade-Off Audit

Use this mode when a study optimizes across multiple environmental or ecosystem-service objectives.

Include:

decision unit;
land-use or management options;
objectives and units;
baseline/reference;
constraints and excluded areas;
model used to generate objective values;
optimization algorithm;
frontier interpretation;
priority maps and consensus;
transition costs, emissions, biodiversity, governance, and realism limits.

Guardrails

Do not pool across species, exposure windows, or outcome constructs without a biological or environmental rationale.
Do not collapse mechanistic, observational, and experimental evidence without design labels.
Do not hide geography, climate zone, tissue type, or measurement platform when they drive heterogeneity.
Do not treat PLS VIP rankings as causal drivers of an environmental outcome; VIP is a model-dependent predictive summary under a chosen component count and scaling.
Do not interpret PLS/VIP without component-selection, cross-validation, coefficient direction, and residual checks.
Do not treat ESR thresholds from PDP, SHAP, ALE, or response curves as causal ecological tipping points without experimental, quasi-experimental, mechanistic, or temporal evidence.
Do not translate model-derived optimal intervals into zoning rules without auditing spatial leakage, resolution, service-model uncertainty, threshold uncertainty, and local feasibility.
Do not treat air-quality food-security counterfactuals as realized production gains without checking crop area, weather, management, prices, imports, storage, and policy implementation.
Do not treat SIF or remote-sensing productivity proxies as yield unless crop-specific validation and conversion logic are explicit.
Do not claim soil biodiversity stabilizes ecosystems under aridity without defining the stability metric, stress gradient, temporal scale, spatial dependence, and possible productivity confounding.
Do not interpret biodiversity-stability moderation as causal unless the design supports causal identification or experimental manipulation.
Do not flatten soil-fauna effects into one "carbon benefit" or "carbon cost" label when stock and flux outcomes move in different directions.
Do not treat coarse-resolution wetland products as scale neutral; audit which small features are omitted.
Do not treat a geospatial environmental map as ground truth; audit observation density, spatial autocorrelation, permafrost or domain masks, predictor time windows, uncertainty layers, and extrapolation zones.
Do not compare cryosphere maps across products or periods without checking target definition, depth convention, permafrost boundary, resolution, map vintage, and uncertainty.
Do not compare methane budgets without checking double-counting between wetlands, lakes, rivers, reservoirs, and inundated vegetation.
Do not turn methane emissions alone into a wetland management recommendation without carbon storage, biodiversity, hydrology, and ecosystem-service trade-offs.
Do not use vote-counting as evidence of effect direction or strength.
Do not treat SHAP, feature importance, PDP, or AutoML performance as causal evidence without a separate identification strategy.
Do not call DML or CATE estimates causal unless treatment timing, confounder adjustment, overlap, and hidden-confounding risks are explicit.
Do not treat DML as a generic prediction method; it estimates causal parameters only under stated identification assumptions.
Do not treat spatial DML as solved by adding coordinates; audit spatial dependence, spillovers, clustered inference, and validation boundaries.
Do not interpret short-term yield gains as soil-carbon sequestration evidence without duration, baseline SOC, depth, and measurement-method checks.
Do not collapse fertilizer form, nutrient rate, baseline nutrient limitation, crop type, and climate into one pooled effect without moderator or subgroup logic.
Do not describe food-system impacts in only one direction when the review question is bidirectional; map both pressure and feedback pathways.
Do not treat global average food footprints as local risk without tracing trade, geography, vulnerability, and production-system context.
Do not treat county-level predicted food waste as directly collectible waste without collection participation, contamination, hauling cost, facility capacity, and policy constraints.
Do not ignore spatial autocorrelation, adjacency leakage, urban-rural imbalance, or high-waste outliers when auditing geospatial ML forecasts.
Do not interpret store counts, restaurant counts, population, income, or social-program access as causal drivers unless the design supports causal inference.
Do not treat random train/test performance in crop trials as proof of generalization to new years, sites, or genotypes; require grouped or out-of-domain validation where the claim needs it.
Do not interpret variable importance, PDPs, or SHAP values causally unless the design supports causal interpretation.
Do not present future yield forecasts without explicit weather/climate inputs, scenario assumptions, and extrapolation checks.
Do not treat a machine-learning scenario projection as causal proof unless the design supports causal interpretation.
Do not present a policy lever, such as recycling rate or intervention coverage, as sufficient without auditing implementation quality and compensating inputs.
Do not treat a Pareto-efficient frontier as a politically feasible pathway without social, economic, governance, biodiversity, and transition-cost constraints.
Do not collapse global objective gains into local welfare claims without regional supply, trade, livelihood, water-access, and justice checks.
Do not generalize an irrigation salinity decision range across regions without soil texture, groundwater depth, climate, crop variety, irrigation method, and long-term salt-balance checks.